From 6abb60c0e16ecde30cb11e109df538f255fa51f1 Mon Sep 17 00:00:00 2001 From: ModelHub XC Date: Tue, 6 Oct 2026 15:47:18 +0800 Subject: [PATCH] =?UTF-8?q?=E5=88=9D=E5=A7=8B=E5=8C=96=E9=A1=B9=E7=9B=AE?= =?UTF-8?q?=EF=BC=8C=E7=94=B1ModelHub=20XC=E7=A4=BE=E5=8C=BA=E6=8F=90?= =?UTF-8?q?=E4=BE=9B=E6=A8=A1=E5=9E=8B?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Model: esmaeil-abedi-dev/verifier-ce-qwen2.5-0.5b Source: Original Platform --- .gitattributes | 36 ++++++++++++++++ README.md | 93 ++++++++++++++++++++++++++++++++++++++++++ chat_template.jinja | 54 ++++++++++++++++++++++++ config.json | 58 ++++++++++++++++++++++++++ generation_config.json | 7 ++++ model.safetensors | 3 ++ tokenizer.json | 3 ++ tokenizer_config.json | 30 ++++++++++++++ 8 files changed, 284 insertions(+) create mode 100644 .gitattributes create mode 100644 README.md create mode 100644 chat_template.jinja create mode 100644 config.json create mode 100644 generation_config.json create mode 100644 model.safetensors create mode 100644 tokenizer.json create mode 100644 tokenizer_config.json diff --git a/.gitattributes b/.gitattributes new file mode 100644 index 0000000..52373fe --- /dev/null +++ b/.gitattributes @@ -0,0 +1,36 @@ +*.7z filter=lfs diff=lfs merge=lfs -text +*.arrow filter=lfs diff=lfs merge=lfs -text +*.bin filter=lfs diff=lfs merge=lfs -text +*.bz2 filter=lfs diff=lfs merge=lfs -text +*.ckpt filter=lfs diff=lfs merge=lfs -text +*.ftz filter=lfs diff=lfs merge=lfs -text +*.gz filter=lfs diff=lfs merge=lfs -text +*.h5 filter=lfs diff=lfs merge=lfs -text +*.joblib filter=lfs diff=lfs merge=lfs -text +*.lfs.* filter=lfs diff=lfs merge=lfs -text +*.mlmodel filter=lfs diff=lfs merge=lfs -text +*.model filter=lfs diff=lfs merge=lfs -text +*.msgpack filter=lfs diff=lfs merge=lfs -text +*.npy filter=lfs diff=lfs merge=lfs -text +*.npz filter=lfs diff=lfs merge=lfs -text +*.onnx filter=lfs diff=lfs merge=lfs -text +*.ot filter=lfs diff=lfs merge=lfs -text +*.parquet filter=lfs diff=lfs merge=lfs -text +*.pb filter=lfs diff=lfs merge=lfs -text +*.pickle filter=lfs diff=lfs merge=lfs -text +*.pkl filter=lfs diff=lfs merge=lfs -text +*.pt filter=lfs diff=lfs merge=lfs -text +*.pth filter=lfs diff=lfs merge=lfs -text +*.rar filter=lfs diff=lfs merge=lfs -text +*.safetensors filter=lfs diff=lfs merge=lfs -text +saved_model/**/* filter=lfs diff=lfs merge=lfs -text +*.tar.* filter=lfs diff=lfs merge=lfs -text +*.tar filter=lfs diff=lfs merge=lfs -text +*.tflite filter=lfs diff=lfs merge=lfs -text +*.tgz filter=lfs diff=lfs merge=lfs -text +*.wasm filter=lfs diff=lfs merge=lfs -text +*.xz filter=lfs diff=lfs merge=lfs -text +*.zip filter=lfs diff=lfs merge=lfs -text +*.zst filter=lfs diff=lfs merge=lfs -text +*tfevents* filter=lfs diff=lfs merge=lfs -text +tokenizer.json filter=lfs diff=lfs merge=lfs -text diff --git a/README.md b/README.md new file mode 100644 index 0000000..b7c2fad --- /dev/null +++ b/README.md @@ -0,0 +1,93 @@ +--- +license: apache-2.0 +base_model: Qwen/Qwen2.5-0.5B +pipeline_tag: text-generation +tags: + - authorization + - ai-agent-safety + - verifier-as-reward + - rlvr +--- + +# verifier-ce-qwen2.5-0.5b + +A **Qwen2.5-0.5B** model fine-tuned to decide **AUTHORIZED / UNAUTHORIZED** +for AI-agent actions over an object-capability delegation chain, using a +**deterministic authorization verifier as the sole training signal** +(verifier cross-entropy — the verdict as a target). + +Companion model to the paper *"Learnable Authorization: A Verifier-as-Reward +Benchmark and Method for AI-Agent Authority."* Code, benchmark, and full +experiment log: https://github.com/esmaeil-abedi-dev/verifier-as-reward + +## What it does + +Given a natural-language description of a root authority, a chain of +delegations (each of which may only narrow the authority it received, with +issue/expiry/revocation times), and a pending action, the model answers +whether the action is authorized. The training target for every example is +the verdict of the deterministic verifier `verify(...)` — never a +hand-assigned or stored label. + +## Training + +- **Objective:** verifier cross-entropy, `-log π(verdict)` (see the paper / + repo for why sampled REINFORCE and exact expected-reward policy gradient + collapse or oscillate at this scale, and why the cross-entropy gradient's + non-saturation is what converges). +- **Data:** synthetic execution traces across 5 domains and 9 scenario + classes, generated and labeled by the verifier; leakage-guarded train / + validation / test splits. +- **Recipe:** class-balanced reward weighting, natural-language prompts, + full-parameter fine-tuning. + +## Evaluation + +Committed held-out test set (`benchmark_test.jsonl`, 80 actions, +natural-language prompts, evaluated once). Three training seeds: + +| seed | accuracy | false-authorize (headline) | false-refuse | +|---|---|---|---| +| 7 | 0.988 | 0.023 | 0.000 | +| 8 | 0.988 | 0.023 | 0.000 | +| **9 (released weights)** | **0.975** | **0.045** | **0.000** | +| mean | 0.983 ± 0.006 | 0.030 ± 0.011 | 0.000 | + +The **weights in this repository are the seed-9 checkpoint** (committed-test +accuracy 0.975; seeds 7/8 reached 0.988). For calibration, on the identical +prompts and metric: claude-sonnet-4.5 0.975, deepseek-r1 0.95, +gemini-2.5-flash 0.875, a lexical heuristic floor 0.80, llama-3.3-70b 0.775, +untrained baseline ~0.47. The only imperfect class is `chain_structure` +(0.90) — the hardest structural-reasoning class. + +**These are in-distribution results** (train/validation/test drawn from one +generator, deduplicated at the decision-context level). They do not by +themselves establish out-of-distribution generalization; see the +repository's domain-hold-out experiment. + +## Intended use and limitations + +- **Research artifact**, not a production authorization system. The verifier + itself is the ground truth; this model approximates it. +- Trained and evaluated on a **synthetic** benchmark from a single verifier; + generalization to real agent logs or to authorization models outside the + attenuated-delegation formalism is not established here. +- See the repository's `EXPERIMENT_LOG.md` for exact metrics (including the + in-distribution vs. out-of-distribution domain-hold-out results) and their + confidence intervals, and `DATASHEET.md` for the benchmark's limitations. + +## Usage + +```python +from transformers import AutoModelForCausalLM, AutoTokenizer +tok = AutoTokenizer.from_pretrained("esmaeil-abedi-dev/verifier-ce-qwen2.5-0.5b") +model = AutoModelForCausalLM.from_pretrained("esmaeil-abedi-dev/verifier-ce-qwen2.5-0.5b") +``` + +Or evaluate it on any benchmark split with the repo's harness (no retraining): + +```bash +PYTHONPATH=. python train_verifier_reward.py \ + --eval-checkpoint esmaeil-abedi-dev/verifier-ce-qwen2.5-0.5b \ + --test-file benchmark_test.jsonl +``` diff --git a/chat_template.jinja b/chat_template.jinja new file mode 100644 index 0000000..28028c0 --- /dev/null +++ b/chat_template.jinja @@ -0,0 +1,54 @@ +{%- if tools %} + {{- '<|im_start|>system\n' }} + {%- if messages[0]['role'] == 'system' %} + {{- messages[0]['content'] }} + {%- else %} + {{- 'You are a helpful assistant.' }} + {%- endif %} + {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within XML tags:\n" }} + {%- for tool in tools %} + {{- "\n" }} + {{- tool | tojson }} + {%- endfor %} + {{- "\n\n\nFor each function call, return a json object with function name and arguments within XML tags:\n\n{\"name\": , \"arguments\": }\n<|im_end|>\n" }} +{%- else %} + {%- if messages[0]['role'] == 'system' %} + {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }} + {%- else %} + {{- '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n' }} + {%- endif %} +{%- endif %} +{%- for message in messages %} + {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %} + {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }} + {%- elif message.role == "assistant" %} + {{- '<|im_start|>' + message.role }} + {%- if message.content %} + {{- '\n' + message.content }} + {%- endif %} + {%- for tool_call in message.tool_calls %} + {%- if tool_call.function is defined %} + {%- set tool_call = tool_call.function %} + {%- endif %} + {{- '\n\n{"name": "' }} + {{- tool_call.name }} + {{- '", "arguments": ' }} + {{- tool_call.arguments | tojson }} + {{- '}\n' }} + {%- endfor %} + {{- '<|im_end|>\n' }} + {%- elif message.role == "tool" %} + {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %} + {{- '<|im_start|>user' }} + {%- endif %} + {{- '\n\n' }} + {{- message.content }} + {{- '\n' }} + {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %} + {{- '<|im_end|>\n' }} + {%- endif %} + {%- endif %} +{%- endfor %} +{%- if add_generation_prompt %} + {{- '<|im_start|>assistant\n' }} +{%- endif %} diff --git a/config.json b/config.json new file mode 100644 index 0000000..af38bbc --- /dev/null +++ b/config.json @@ -0,0 +1,58 @@ +{ + "architectures": [ + "Qwen2ForCausalLM" + ], + "attention_dropout": 0.0, + "bos_token_id": 151643, + "dtype": "float16", + "eos_token_id": 151643, + "hidden_act": "silu", + "hidden_size": 896, + "initializer_range": 0.02, + "intermediate_size": 4864, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 32768, + "max_window_layers": 24, + "model_type": "qwen2", + "num_attention_heads": 14, + "num_hidden_layers": 24, + "num_key_value_heads": 2, + "pad_token_id": null, + "rms_norm_eps": 1e-06, + "rope_parameters": { + "rope_theta": 1000000.0, + "rope_type": "default" + }, + "sliding_window": null, + "tie_word_embeddings": true, + "transformers_version": "5.13.1", + "use_cache": true, + "use_mrope": false, + "use_sliding_window": false, + "vocab_size": 151936 +} diff --git a/generation_config.json b/generation_config.json new file mode 100644 index 0000000..dd9f05c --- /dev/null +++ b/generation_config.json @@ -0,0 +1,7 @@ +{ + "bos_token_id": 151643, + "do_sample": false, + "eos_token_id": 151643, + "max_new_tokens": 2048, + "transformers_version": "5.13.1" +} diff --git a/model.safetensors b/model.safetensors new file mode 100644 index 0000000..62a67a5 --- /dev/null +++ b/model.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eefae33fd8974308a190a03b3329e09b8b0327f77aff780a339588ddc7174612 +size 988097536 diff --git a/tokenizer.json b/tokenizer.json new file mode 100644 index 0000000..34510ff --- /dev/null +++ b/tokenizer.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fd169731d2cbde95e10bf356d66d5997fd885dd8dbb6fb4684da3f23b2585d8 +size 11421892 diff --git a/tokenizer_config.json b/tokenizer_config.json new file mode 100644 index 0000000..1127e12 --- /dev/null +++ b/tokenizer_config.json @@ -0,0 +1,30 @@ +{ + "add_prefix_space": false, + "backend": "tokenizers", + "bos_token": null, + "clean_up_tokenization_spaces": false, + "eos_token": "<|endoftext|>", + "errors": "replace", + "extra_special_tokens": [ + "<|im_start|>", + "<|im_end|>", + "<|object_ref_start|>", + "<|object_ref_end|>", + "<|box_start|>", + "<|box_end|>", + "<|quad_start|>", + "<|quad_end|>", + "<|vision_start|>", + "<|vision_end|>", + "<|vision_pad|>", + "<|image_pad|>", + "<|video_pad|>" + ], + "is_local": true, + "local_files_only": false, + "model_max_length": 131072, + "pad_token": "<|endoftext|>", + "split_special_tokens": false, + "tokenizer_class": "Qwen2Tokenizer", + "unk_token": null +}